WO2022053001A1

WO2022053001A1 - Weld seam internal defect intelligent detection device and method, and medium

Info

Publication number: WO2022053001A1
Application number: PCT/CN2021/117548
Authority: WO
Inventors: 刘骁佳; 戴铮; 刘欢; 周鹏飞; 王飞; 王英伟; 洪海波; 危荃
Original assignee: 上海航天精密机械研究所
Priority date: 2020-09-10
Filing date: 2021-09-10
Publication date: 2022-03-17
Also published as: CN112083017A

Abstract

A weld seam internal defect intelligent detection device and method, and a detection medium. An X-ray is emitted by an X-ray tube to perform radiographic inspection, an optical image is achieved by means of an optical path simulation and control unit, and a digital image containing weld seam internal quality information is obtained on an imaging plate. The generated image is automatically uploaded to a cloud platform server, undergoes intelligent pre-processing using digital image processing technologies and deep leaning neural network algorithms, and is analyzed to determine defect existence, defect position, defect type and defect rating. Thus the quality detection of the image presented is achieved. The weld seam detection of a complex structure is accurately controlled during the detection process. Manual assessment is replaced by intelligent defect recognition during the image evaluation process, the length of manual detection is effectively shortened, and the efficiency of weld seam quality detection is improved, while the accuracy of defect recognition is guaranteed.

Description

Intelligent detection device, method and medium for welding internal defects

technical field

The invention relates to the technical field of intelligent detection, in particular, to an intelligent detection device, method and medium for internal defects of welds.

Background technique

At present, in the industrial field, the detection of internal weld defects is an important part of the quality inspection of the production process. Usually, digital radiography technology is used to film the internal information of the weld, and then experienced workers interpret the internal defects of the weld according to the film. However, the defect evaluation of digital radiographic images requires qualified skilled workers to complete, and it takes a lot of time and money to train skilled workers. Moreover, a large number of film evaluations can easily lead to fatigue of skilled workers, which can easily lead to misjudgment and missed judgment. etc. Human error. At the same time, with the development of productivity and the increase of output, higher requirements have been put forward for the efficiency of film evaluation. With the development of artificial intelligence technology, the deep learning neural network algorithm based on digital images has assumed more and more important tasks in the field of non-destructive testing.

In the prior art, the inspection personnel formulate the inspection process to realize the welding seam inspection. The layout of the inspection optical path depends heavily on the inspection personnel's knowledge reserve and experience, and the inspection image evaluation depends on the inspector. not tall.

Patent document CN107748200B (application number: CN201710712559.0) discloses a piezoelectric array flexible sensor and a detection method for detecting weld seam defects based on characteristic guided waves. The sensor is composed of a plurality of piezoelectric units, which are arranged on a flexible substrate; each piezoelectric unit is covered with a damping block, surrounded by sound-absorbing materials, and the sensor shell is packaged with a flexible protective film; all piezoelectric units are connected in series with a logic After switch and delay, connect in parallel to the positive bus. Piezoelectric units are divided into three categories, which excite three different modes of guided waves respectively, and have different degrees of sensitivity to different types of defects to achieve complementarity. The sensor selects different piezoelectric array detection methods through logic switches, and adapts to the welding seam detection requirements of different curvature surfaces and structures; the delay device adjusts the excitation time difference to realize the synthesis and focusing of the sound beam.

SUMMARY OF THE INVENTION

Aiming at the defects in the prior art, the purpose of the present invention is to provide an intelligent detection device, method and medium for internal defects of welds.

An intelligent detection device for welding internal defects provided according to the present invention includes: an X-ray tube, an imaging board, an optical path simulation and control unit, and a cloud platform server;

The optical path simulation and control unit selects the transillumination optical path according to the shape of the weld, and adjusts the positions of the X-ray tube and the imaging plate according to the selected transillumination optical path, so that the weld is located at the preset position between the X-ray tube and the imaging plate; X After the ray transmits the weld, it is imaged on the imaging plate to obtain the X-ray image, that is, the original grayscale image of the weld, and the image is pushed to the cloud platform server.

Preferably, the preset position refers to:

The imaging plate is located in the focal plane where the focal point of the radiation emitted by the X-ray tube is located.

According to an intelligent detection method for welding internal defects based on the above-mentioned intelligent detection device for welding internal defects provided by the present invention, the optical path simulation and control unit selects the optimal transillumination optical path according to the shape of the weld, and according to the optimal penetration Illuminate the light path, adjust the position of the X-ray tube and the imaging plate, so that the welding seam is at the preset position between the X-ray tube and the imaging plate; stitch the original grayscale image, and push the image to the cloud platform server;

The cloud platform server preprocesses the original grayscale image of the weld, so that the contrast of the original grayscale image meets the requirements, and the preprocessed image is obtained; the preprocessed image is input into the convolutional neural network model CNN for defect screening, Determine whether the preprocessed image contains defects;

Input the image judged to contain defects into the convolutional neural network model RCNN with the region of interest, locate the defect in the defective image, and mark the minimum block diagram of the defect location to realize defect localization;

Extract the image in the minimum block diagram after defect location, input the deep learning neural network model DNN, classify the defect, and get the type of defect.

Preferably, the original grayscale image of the weld is preprocessed, and the preprocessing method includes: image enhancement, non-linear image change, filtering and noise reduction, and image sharpening;

Contrast requirements of the original grayscale image, specifically: the grayscale difference between four adjacent pixels of one pixel meets the minimum distinction requirement of the detection threshold;

Defects, including pores, slag inclusions, shrinkage porosity, cracks, and segregation.

Preferably, the convolutional neural network model CNN is pre-trained, and the specific training method is as follows:

(1) Establish a labelled training image set, the training image is the preprocessed image of the original grayscale image of the weld, and the image is pre-marked with or without defects to form a label; the labels are divided into defective and non-defective labels;

(2) Input the images in the training image set into the untrained convolutional neural network model CNN one by one. According to the pre-set structure, objective function and model parameter optimization method of the convolutional neural network model CNN, the convolutional neural network model CNN is trained to obtain a trained convolutional neural network model CNN, that is, the parameters in the convolutional neural network model CNN have been optimized.

Preferably, the trained convolutional neural network model CNN can output a label, and judge whether the weld is defective according to the label, so as to realize defect screening.

The minimum block diagram of the location of the defect refers to: a rectangular frame composed of a set of parallel lines tangent to the longest diameter of the defect and a set of parallel lines perpendicular to it.

Preferably, the convolutional neural network model RCNN with the region of interest needs to be pre-trained, and the training method is as follows:

(1) Establish a training image set with a minimal block diagram of defects;

(2) Input the images in the training image set with the minimum defect block diagram to the initial convolutional neural network model RCNN one by one, and set the structure, objective function, and model parameters of the convolutional neural network model RCNN with the region of interest in advance. The optimization method is to train the convolutional neural network model RCNN with the region of interest, and obtain the trained RCNN model with the region of interest, that is, the parameters in the convolutional neural network model RCNN with the region of interest have been processed. optimization.

Preferably, the trained convolutional neural network model RCNN with a region of interest can output the minimum block diagram of the location of the defect, and realize the defect location according to the minimum block diagram of the location of the defect.

Preferably, the deep learning neural network model DNN needs to be pre-trained, and the training method is as follows:

(1) Establish a training image set with a defect minimum block diagram with a defect type label; the defect type label of each training image with a defect minimum block diagram corresponds to a defect type;

(2) Input the images in the training image set with the defect minimum block diagram with the defect type label into the initial deep learning neural network model DNN one by one, and optimize the method according to the structure, objective function and model parameter of the deep learning neural network model DNN in advance. , train the deep learning neural network model DNN to obtain a trained deep learning neural network model DNN, that is, the parameters in the deep learning neural network model DNN have been optimized.

Preferably, when the computer program is executed by the processor, the steps of any one of the above-mentioned methods for intelligently detecting internal defects of a weld seam are implemented.

Compared with the prior art, the present invention has the following beneficial effects:

(1) The present invention realizes the confirmation of the optimal detection optical path by adopting the optical path simulation and control module, instead of the optical path arrangement determined by the testing personnel relying on experience, to obtain the optimal image and improve the image accuracy and sensitivity.

(2) The present invention uses the cloud platform server to perform intelligent analysis of digital ray images, and through cloud services, can effectively collect and summarize the gradual quality information and internal defects of different products in the inspection process, and improve data support for design and production process optimization. , and ultimately improve the quality of the product

(3) The present invention analyzes digital ray images and intelligently interprets the results through the computer deep learning method, partially or completely replaces the manual film evaluation process, effectively reduces the manual detection time, and avoids human errors under the premise of ensuring the accuracy of defect recognition. , to improve the work efficiency of weld quality inspection.

(4) The present invention uses the convolutional neural network model CNN to screen digital ray image defects. This method can effectively identify the macroscopic information and microscopic information in the image, and can upgrade and optimize the model with the accumulation of data, so as to iterate for the defect screening algorithm. Provide support.

(5) The present invention uses the convolutional neural network model RCNN with the region of interest to locate defects in digital ray images. This method can effectively identify the minimum block diagram of defects and provide support for subsequent defect size measurement for defect rating.

(6) Real-time uploading and processing of detection data through cloud platform server, creating conditions for the digitization of the entire production process

Description of drawings

Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments with reference to the following drawings:

Fig. 1 is the schematic block diagram of the intelligent defect detection method of welding seam digital radiographic image according to the present invention;

Fig. 2 is a schematic diagram of an intelligent detection device for internal defects of welds according to the present invention;

FIG. 3 is a schematic diagram of a welding seam digital radiographic image defect screening algorithm according to the present invention;

FIG. 4 is a schematic diagram of the defect location algorithm of the digital radiographic image of the weld according to the present invention.

detailed description

The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the inventive concept. These all belong to the protection scope of the present invention.

The present invention will be described in more detail below through examples.

Example 1:

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

The invention is mainly applied to the welding seam inspection of complex structural parts, effectively improves inspection quality and inspection consistency, improves the accuracy and efficiency of inspection image evaluation, and finally improves the welding seam quality. It provides the basis for the digitalization and intelligent application of production processes.

Fig. 1 is a schematic block diagram of the intelligent defect detection method for welding seam digital radiographic image according to the present invention. Defect 1, 7 - Defect 2, 8 - Defect 3, 9 - Qualified, 10 - Rework Repair, 11 - Concession Acceptance, 12 - Scrap. Fig. 2 is a schematic diagram of the intelligent detection device for welding internal defects of the present invention. In Fig. 2, ①-optical path simulation and control unit, ②-X-ray tube, ③-imaging board, and ④-cloud platform. Fig. 3 is a schematic diagram of the defect screening algorithm of the digital radiographic image of the weld according to the present invention. In Fig. 3, a-convolutional layer, b-downsampling layer, c-convolutional layer, d-downsampling layer, e-full connection layer, f - Output layer (full connection + Softmax activation). Fig. 4 is a schematic diagram of the welding seam digital radiographic image defect location algorithm of the present invention. In Fig. 4, A-extracts candidate frames, B-CNN extracts feature maps, C-ROI pooling, D-classification, and 5-regression.

An intelligent detection device for welding internal defects of the present invention includes: an X-ray tube, an imaging board, an optical path simulation and control unit, and a cloud platform server; as shown in FIG. 2 ;

The optical path simulation and control unit can optimize the optimal transillumination optical path through the built-in simulation module according to the topographical features of the weld or the 3D model, that is, the X-rays are emitted from the ray tube, perpendicular to the surface of the weld or its sectional plane to transmit the weld to the imaging When it is impossible to inject from the vertical surface due to the special structure of the welding seam, ensure the shortest path of the radiation transmission welding seam; control the X-ray tube and the imaging plate according to the optimal transmission pipeline, and adjust their positions so that the welding seam is located in the X-ray tube. The best position between it and the imaging plate, that is, the imaging plate is located in the focal plane where the focal point of the X-ray tube emits rays; the optical path simulation and control unit controls the X-ray tube to emit X-rays. Imaging, the optical path simulation and control unit controls the imaging board to obtain the X-ray image, that is, the original grayscale image of the weld; the optical path simulation and control unit pushes the original grayscale image of the weld to the cloud platform server;

The cloud platform server preprocesses the original grayscale image of the weld, so that the contrast of the original grayscale image meets the requirements, and the preprocessed image is obtained; the preprocessed image is input into the convolutional neural network model CNN for defect screening, Determine whether the preprocessed image contains defects. The cloud platform server can package and integrate defect intelligent detection methods, and can also be deployed on the Internet to output cloud services; it can effectively collect and summarize different products in the inspection process, and form weld digital ray image data sets of different products, which is the basis for defect intelligence. The development of detection methods provides the basis.

The image judged to contain defects is input into the convolutional neural network model RCNN with the region of interest, the defects in the defective images are located, and the minimum block diagram of the location of the defect is marked to realize the defect localization.

The imaging plate is located at the focal point of the X-ray tube radiation.

Preprocess the original grayscale image of the weld as follows:

For the original grayscale image with sharp changes in thickness gradient, the preprocessing methods include but are not limited to: image enhancement, image nonlinear change, filter noise reduction, image sharpening, and the like.

For the original grayscale image with rough surface, the preprocessing methods include but are not limited to: image enhancement, filtering and noise reduction, and image smoothing.

For the original grayscale image of the edge of the weld, the preprocessing methods include but are not limited to: image enhancement, non-linear image change, filtering and noise reduction, wavelet transform, and the like.

In actual image preprocessing, different combinations and parameters of preprocessing methods will be tried, and the combination and parameters of preprocessing methods with better effect will be selected from among them.

The contrast requirement of the original grayscale image is preferably as follows: the grayscale difference of four adjacent pixel points of one pixel point meets the minimum discrimination threshold requirement of the detection threshold.

Defects preferably include porosity, slag inclusion, and incomplete penetration.

The convolutional neural network model CNN is pre-trained, and the preferred training method is as follows:

①Establish a labelled training image set, the training image is the preprocessed image of the original grayscale image of the weld, and mark the image with or without defects in advance to form a label; the label is divided into defective and non-defective labels;

② Input the images in the training image set into the untrained convolutional neural network model CNN one by one, that is, initialize the model parameters with random functions on the basis of the VGG network model structure, and then define the error between the model output and the input as the objective function, The model parameters are updated by the method of error back propagation, and the minimum value of the objective function is obtained under the parameter update by the method of gradient descent. The above process is used to train the convolutional neural network model CNN, as shown in Figure 3, after the image is input , the feature map is obtained by convolution, the feature map is obtained by down-sampling, and the process of convolution and down-sampling is repeated, and the final feature map is input into the fully connected layer, and the screening result is output through the output layer. After training, the trained convolutional neural network model CNN is obtained, that is, the parameters in the convolutional neural network model CNN have been optimized. In the actual training process, the structure, objective function, and model parameters of the convolutional neural network model CNN can be optimized in various combinations for training and testing. For example, in addition to the VGG network, the structure of CNN can also choose network structures such as AlexNet and ResNet; the objective function In addition to the error between output and input, parameters such as learning rate and regular term can also be added to form a relatively complex objective function; in addition to the gradient descent method, model parameter optimization can also choose stochastic gradient descent, momentum gradient descent, and mini-batch gradient descent. When training, you can choose various methods to combine, and choose a combination with better effect, such as high accuracy, low false detection rate, fast training speed, etc., as parameters for convolutional neural network model CNN training.

The trained convolutional neural network model CNN can output the label, judge whether the weld is defective according to the label, and realize defect screening.

The convolutional neural network model RCNN with regions of interest needs to be pre-trained. The preferred training method is as follows:

①Establish a training image set with the smallest block diagram of defects; the training images in the training image set contain all possible defects of the weld, each training image contains at least one defect, and all possible defects, including: pores, slag inclusions , not fully welded.

② Input the images in the training image set with the smallest defect block diagram one by one into the untrained convolutional neural network model RCNN with regions of interest, that is, use random functions to initialize the model parameters on the basis of the YOLO network model structure, and then The error between the output and input of the model is defined as the objective function, the model parameters are updated by the method of error back propagation, and the gradient descent method is used to find the parameter update and drop to obtain the minimum value of the objective function. The convolutional neural network model RCNN is trained, as shown in Figure 4, the candidate frame is extracted from the input image, the candidate frame is mapped to the feature map ROI by the method of extracting the feature map by CNN, and then the ROI is adjusted to a fixed size by ROI pooling, Classification and regression are performed after obtaining ROI features. After training, the trained convolutional neural network model RCNN with the region of interest is obtained, that is, the parameters in the convolutional neural network model RCNN with the region of interest have been optimized. In the actual training process, the structure, objective function, and model parameters of the convolutional neural network model RCNN can be combined for training and testing. For example, in addition to the YOLO network, the structure of RCNN can also choose SDD and other network structures; In addition to the input error, parameters such as learning rate and regular term can also be added to form a relatively complex objective function; in addition to the gradient descent method for model parameter optimization, stochastic gradient descent, momentum gradient descent, and mini-batch gradient descent can also be selected. During training, you can choose various methods to combine, and choose a combination with good effect, such as high accuracy, low false detection rate, fast training speed, etc., as parameters to train the convolutional neural network model RCNN.

The trained convolutional neural network model RCNN with a region of interest is preferably able to output the smallest block diagram of the location of the defect, and realize the defect location according to the smallest block diagram of the location of the defect.

The deep learning neural network model DNN needs to be pre-trained. The preferred training method is as follows:

①Establish a training image set with a defect minimum block diagram with a defect type label; the defect type label of each training image with a defect minimum block diagram corresponds to a defect type;

② Input the images in the training image set with the defect type label with the smallest defect block diagram one by one into the untrained deep learning neural network model DNN, that is, on the basis of the AlexNet network model structure, use a random function to initialize the model parameters, and then use a random function to initialize the model parameters. The error between the output and input of the model is defined as the objective function, and the model parameters are updated by the method of error back propagation, and the gradient descent method is used to find the parameter update and drop to obtain the minimum value of the objective function. After training, the trained deep learning neural network model DNN is obtained, that is, the parameters in the deep learning neural network model DNN have been optimized. In the actual training process, the structure, objective function, and model parameters of the convolutional neural network model DNN can be combined for training and testing. For example, in addition to the AlexNet network, the structure of the DNN can also choose VGG, ResNet and other network structures; In addition to the error between output and input, parameters such as learning rate and regular term can also be added to form a relatively complex objective function; in addition to gradient descent, model parameter optimization can also choose stochastic gradient descent, momentum gradient descent, and mini-batch gradient descent. When training, you can choose various methods to combine, and choose a combination with good effect, such as high accuracy, low false detection rate, fast training speed, etc., as parameters for DNN training of convolutional neural network model.

The deep learning neural network model DNN can output the label with the defect type, and realize the defect classification according to the label with the defect type.

After the defects are classified, the classified defects can also be graded to determine whether their defect grades meet the pre-input patterns of different grades of the same type of defects, and output the detection results;

The present invention performs preprocessing to achieve a further solution for improving defect contrast as follows: introducing an optimization algorithm to find a preprocessing combination that maximizes defect contrast.

The present invention sets the contrast requirement of the original grayscale image, and a further solution for realizing the improvement of the image contrast is: adjusting the X-ray irradiation intensity according to the size and material of the weld to be inspected.

In the present invention, the convolutional neural network model CNN is trained, and the further scheme for realizing the improvement of defect screening is: selecting the structure of the residual network ResNet as the structure of the volume and the neural network model CNN for training.

According to the present invention, the convolutional neural network model RCNN with the region of interest is trained, and the further scheme for realizing the improvement of defect location is: selecting the structure of the Faster R-CNN model as the structure of the volume with the region of interest and the neural network model RCNN to carry out the training. Training.

The deep learning neural network model DNN of the present invention is trained, and the further scheme for realizing the improvement of defect classification is: optimizing the learning rate and the regular term part in the objective function of the deep learning neural network model DNN

After the defects are classified, the present invention can also carry out defect ratings for the classified defects, judge whether the defect grades meet the inspection standards, and output the inspection results. The preferred solution is:

After the defects are classified, the DNN network is used to classify the classified defects, and the defects are rated. Specifically, the images to be rated are input into the trained network, and the labels representing the defect levels are output. The judgment is based on evaluating whether it meets the previous DNN network. Patterns of different levels of the same defect entered by the trainer.

A further scheme for realizing the improvement of the device index by the present invention is: a further scheme for realizing the reduction of the missed judgment rate of intelligent film evaluation is to write the missed judgment rate into the objective function for optimization.

In the welding seam inspection process, the present invention overcomes the problem of filming the internal information of the welding seam by the commonly used radiographic imaging technology, and then interpreting the internal defects of the welding seam by experienced workers according to the film. Moreover, during the quality inspection period, the labor cost is high, the experienced workers are scarce, and the training cycle of skilled workers with inspection qualifications is long. With the increase of inspection work intensity, manual inspection is often accompanied by low inspection efficiency, misjudgment and leakage of inspection results. Judgment and other human errors, which seriously affect the product quality inspection work.

Taking the girth weld of the cabin as an example, the key process of the intelligent detection device and method for weld defects proposed by the present invention is as follows: (1) As for the ray source inside the cylindrical cabin, the X-rays are emitted from the ray source and are perpendicular to the weld cabin. The cut surface of the body surface is injected to form the best transillumination light path, which is received by the imaging plate placed outside the cabin and located on the ray focal plane to form the original grayscale image of the digital ray of the girth weld of the cabin, which is pushed to the cloud platform server. (2) The cloud platform server preprocesses the original grayscale image of the digital ray of the cabin girth weld, so that the contrast of the original grayscale image meets the requirements, and the preprocessed image is obtained; (3) The preprocessed image is input into the volume The integrated neural network model CNN is used to screen defects and determine whether the preprocessed images contain defects. The convolutional neural network model CNN needs to be pre-trained. On the basis of the VGG network model structure, the model parameters are initialized with random functions to obtain an untrained convolutional neural network model CNN. The error between the model output and input is defined as the objective function, and the error is used. The method of backpropagation updates the model parameters, and the method of gradient descent is used to find the parameter update and drop to obtain the minimum value of the objective function, and complete the training of the convolutional neural network model CNN. The trained convolutional neural network model CNN has the ability to screen defects and can screen the input images for defects. (4) Input the images judged to contain defects into the convolutional neural network model RCNN with the region of interest, locate the defects in the defective images, and mark the minimum block diagram of the location of the defects, so as to realize the defect localization. The convolutional neural network model RCNN with a region of interest needs to be pre-trained. On the basis of the YOLO network model structure, a random function is used to initialize the model parameters to obtain an untrained convolutional neural network model RCNN with a region of interest. The error between the output and the input is defined as the objective function, the model parameters are updated by the method of error back propagation, and the gradient descent method is used to find the parameter update and drop to obtain the minimum value of the objective function, and complete the convolutional neural network with the region of interest. Training of the model RCNN. The trained convolutional neural network model RCNN with regions of interest has the capability of defect localization, and can mark the smallest block diagram of the location of the defect to realize defect localization. (5) Extract the image in the minimum block diagram after defect location, input the deep learning neural network model DNN, classify the defect, and obtain the type of defect. The deep learning neural network model DNN needs to be pre-trained. On the basis of the AlexNet network model structure, a random function is used to initialize the model parameters to obtain an untrained deep learning neural network model DNN. The error between the model output and the input is defined as the objective function, using the error The method of back propagation is used to update the model parameters, and the method of gradient descent is used to find the parameter update and lower to obtain the minimum value of the objective function, and complete the training of the deep learning neural network model DNN. The trained deep learning neural network model DNN has the ability to classify defects, and can classify the specific categories of defect images (porosity, slag inclusion, and incomplete penetration) in the input minimum block diagram.

In the present invention, the evaluation efficiency test is carried out, 500 images are randomly selected from the digital radiographic images of the welding seam, and 500 images are handed over to experienced film critics for evaluation and the intelligent defect detection device and method mentioned in the present invention. The total time for manual and machine evaluation is calculated, and the evaluation time for each image is calculated. The test results are that the manual filming time is 28.6 seconds per sheet, and the machine filming time is 14.3 seconds per sheet. The defect intelligent detection device and method mentioned in the present invention improves the efficiency by about 100% compared with the traditional manual evaluation method.

It can be seen from the comparison that in the present invention, the traditional manual detection method is replaced by the detection method of computer combined with deep learning neural network artificial intelligence algorithm, and the detection knowledge of experienced inspection technicians is converted into machine language that can be quantified by computer. The digital image collected by the inspection equipment is analyzed by new artificial intelligence technology, and the defects in the image are located, identified, rated, and automatically and intelligently interpret the output results. , to avoid human error, improve work efficiency and reduce labor costs. And when the amount of detection data continues to accumulate, the detection accuracy rate will also increase.

Example 2:

The technical problem solved by the present invention is: to overcome the deficiencies of the prior art, to provide an intelligent detection device and method for internal defects of the weld, which is applied to the quality inspection task of the weld. X-ray image analysis, intelligent interpretation of the results, effectively reduce the time of manual inspection, avoid errors caused by human operation under the premise of ensuring the accuracy of defect identification, and improve the efficiency of weld quality inspection.

The technical scheme solved by the present invention is: an intelligent detection device and method for welding internal defects, comprising: an X-ray tube, an imaging board, an optical path simulation and control unit, and a cloud platform server;

The optical path simulation and control unit can select the optimal transillumination optical path according to the shape or 3D model of the weld, and can adjust the position of the X-ray tube and the imaging plate according to the optimal transillumination optical path, so that the weld is located between the X-ray tube and the imaging plate. The optimal position between the imaging plates; the optical path simulation and control unit controls the X-ray tube to emit X-rays, and after the X-rays transmit the welding seam, they are imaged on the imaging plate to obtain the X-ray image, that is, the original grayscale image of the welding seam; The original grayscale image of the seam is pushed to the cloud platform server;

The cloud platform server preprocesses the original grayscale image of the weld so that the contrast of the original grayscale image meets the requirements, that is, the preprocessed image is obtained; the preprocessed image is input into the convolutional neural network model CNN for defect screening , to determine whether the preprocessed image contains defects.

Preferably, the imaging plate is to be located at the focal point of the X-ray tube radiation.

Preferably, the original grayscale image of the weld is preprocessed, as follows:

For the original grayscale image with sharp changes in thickness gradient, the preprocessing methods include but are not limited to: image enhancement, image nonlinear change, filter noise reduction, and image sharpening.

For the original grayscale image of the edge of the weld, the preprocessing methods include but are not limited to: image enhancement, non-linear image change, filtering and noise reduction, and wavelet transform.

Preferably, the contrast requirement of the original grayscale image is specifically: the grayscale difference of four adjacent pixels of one pixel meets the minimum discrimination requirement of the detection threshold.

Preferably, the defects include porosity, slag inclusion, and incomplete penetration.

(2) Input the images in the training image set into the untrained convolutional neural network model CNN one by one. According to the structure, objective function and model parameter optimization method of the pre-set convolutional neural network model CNN, the convolutional neural network is optimized. The model CNN is trained to obtain a trained convolutional neural network model CNN, that is, the parameters in the convolutional neural network model CNN have been optimized.

Preferably, the minimum block diagram of the location of the defect refers to: a rectangular frame composed of a group of parallel lines tangent to the longest diameter of the defect and a group of parallel lines perpendicular to it (the part that can be tangent to the outermost edge of the defect).

(1) Establish a training image set with a minimal block diagram of defects;

(2) Input the images in the training image set with the minimum defect block diagram one by one into the untrained convolutional neural network model RCNN with the region of interest, and set the convolutional neural network model with the region of interest according to the preset The structure, objective function, and model parameter optimization method of RCNN, train the convolutional neural network model RCNN with the region of interest, and obtain the trained convolutional neural network model RCNN with the region of interest, that is, the model with the region of interest has been trained. The parameters in the convolutional neural network model RCNN of the region of interest are optimized.

(2) Input the images in the training image set with the defect minimum block diagram with the defect type label one by one into the untrained deep learning neural network model DNN, and set the structure, objective function, In the model parameter optimization method, the deep learning neural network model DNN is trained to obtain a trained deep learning neural network model DNN, that is, the parameters in the deep learning neural network model DNN have been optimized.

Preferably, the deep learning neural network model DNN can output labels with defect types, and realize defect classification according to the labels with defect types.

Preferably, after the defects are classified, a defect rating can be performed on the classified defects to determine whether the defect grades satisfy the different level patterns of the same type of defects input by the pre-DNN network training, and output the detection results.

In the description of this application, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", The orientation or positional relationship indicated by "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present application and simplifying the description, rather than indicating or implying the indicated device. Or elements must have a particular orientation, be constructed and operate in a particular orientation, and therefore should not be construed as a limitation of the present application.

Those skilled in the art know that, in addition to implementing the system, device and each module provided by the present invention in the form of pure computer readable program code, the system, device and each module provided by the present invention can be completely implemented by logically programming the method steps. The same program is implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, and embedded microcontrollers, among others. Therefore, the system, device and each module provided by the present invention can be regarded as a kind of hardware component, and the modules used for realizing various programs included in it can also be regarded as the structure in the hardware component; A module for realizing various functions can be regarded as either a software program for realizing a method or a structure within a hardware component.

Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above-mentioned specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essential content of the present invention. The embodiments of the present application and features in the embodiments may be combined with each other arbitrarily, provided that there is no conflict.

Claims

An intelligent detection device for internal defects of welds, which is characterized by comprising: an X-ray tube, an imaging board, an optical path simulation and control unit, and a cloud platform server;

The optical path simulation and control unit selects the transillumination optical path according to the shape of the weld, and adjusts the positions of the X-ray tube and the imaging plate according to the selected transillumination optical path, so that the weld is located at the preset position between the X-ray tube and the imaging plate; X After the ray transmits the weld, it is imaged on the imaging plate to obtain the X-ray image, that is, the original grayscale image of the weld, and the image is pushed to the cloud platform server.
The intelligent detection device and method for welding internal defects according to claim 1, wherein the preset position refers to:

The imaging plate is located in the focal plane where the focal point of the radiation emitted by the X-ray tube is located.
An intelligent detection method for welding internal defects based on the intelligent detection device for welding internal defects according to claim 1, characterized in that the optical path simulation and control unit selects the optimal transillumination optical path according to the shape of the welding seam, and according to the optimal Transilluminate the light path, adjust the position of the X-ray tube and the imaging plate, so that the weld is located at the preset position between the X-ray tube and the imaging plate; after the X-ray transmits the weld, it is imaged on the imaging plate to obtain the X-ray image, that is The original grayscale image of the weld, push the image to the cloud platform server;

The cloud platform server preprocesses the original grayscale image of the weld, so that the contrast of the original grayscale image meets the requirements, and the preprocessed image is obtained; the preprocessed image is input into the convolutional neural network model CNN for defect screening, Determine whether the preprocessed image contains defects;

Input the image judged to contain defects into the convolutional neural network model RCNN with the region of interest, locate the defect in the defective image, and mark the minimum block diagram of the defect location to realize defect localization;

Extract the image in the minimum block diagram after defect location, input the deep learning neural network model DNN, classify the defect, and get the type of defect.
An intelligent detection method for welding internal defects according to claim 3, wherein the original grayscale image of the welding seam is preprocessed, and the preprocessing methods include: image enhancement, non-linear image change, filtering Noise reduction and image sharpening;

Contrast requirements of the original grayscale image, specifically: the grayscale difference between four adjacent pixels of one pixel meets the minimum distinction requirement of the detection threshold;

Defects, including pores, slag inclusions, shrinkage porosity, cracks, and segregation.
A kind of intelligent detection method of weld internal defect according to claim 3, is characterized in that, described convolutional neural network model CNN is through pre-training, and concrete training mode is as follows:

(1) Establish a labelled training image set, the training image is the preprocessed image of the original grayscale image of the weld, and the image is pre-marked with or without defects to form a label; the labels are divided into defective and non-defective labels;

(2) Input the images in the training image set into the untrained convolutional neural network model CNN one by one. According to the pre-set structure, objective function and model parameter optimization method of the convolutional neural network model CNN, the convolutional neural network model CNN is trained to obtain a trained convolutional neural network model CNN, that is, the parameters in the convolutional neural network model CNN have been optimized.
The intelligent detection method for welding seam internal defects according to claim 5, wherein the trained convolutional neural network model CNN can output labels, and judge whether the welding seam is defective according to the labels, so as to realize defect screening.

The minimum block diagram of the location of the defect refers to: a rectangular frame composed of a set of parallel lines tangent to the longest diameter of the defect and a set of parallel lines perpendicular to it.
A method for intelligent detection of internal defects in welds according to claim 3, wherein the convolutional neural network model RCNN with a region of interest needs to be pre-trained, and the training method is as follows:

(1) Establish a training image set with a minimal block diagram of defects;

(2) Input the images in the training image set with the minimum defect block diagram to the initial convolutional neural network model RCNN one by one, and set the structure, objective function, and model parameters of the convolutional neural network model RCNN with the region of interest in advance. The optimization method is to train the convolutional neural network model RCNN with the region of interest, and obtain the trained RCNN model with the region of interest, that is, the parameters in the convolutional neural network model RCNN with the region of interest have been processed. optimization.
The intelligent detection method for internal defects in welds according to claim 7, wherein the trained convolutional neural network model with a region of interest (RCNN) can output the minimum block diagram of the location of the defect, according to the location of the defect. A minimal block diagram of the location, enabling defect localization.
The intelligent detection method for welding internal defects according to claim 3, wherein the deep learning neural network model DNN needs to be pre-trained, and the training method is as follows:

(1) Establish a training image set with a defect minimum block diagram with a defect type label; the defect type label of each training image with a defect minimum block diagram corresponds to a defect type;

(2) Input the images in the training image set with the defect minimum block diagram with the defect type label one by one into the initial deep learning neural network model DNN, and optimize the method according to the structure, objective function and model parameter of the deep learning neural network model DNN in advance. , train the deep learning neural network model DNN to obtain a trained deep learning neural network model DNN, that is, the parameters in the deep learning neural network model DNN have been optimized.
A computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, the steps of the intelligent detection method for welding internal defects according to any one of claims 3 to 9 are implemented.